NADI 2021: The Second Nuanced Arabic Dialect Identification Shared Task - Computational Approaches to Modeling ...
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
NADI 2021: The Second Nuanced Arabic Dialect Identification Shared Task Muhammad Abdul-Mageed, Chiyu Zhang, AbdelRahim Elmadany, Houda Bouamor,† Nizar Habash‡ The University of British Columbia, Vancouver, Canada † Carnegie Mellon University in Qatar, Qatar ‡ New York University Abu Dhabi, UAE {muhammad.mageed, a.elmadany}@ubc.ca chiyuzh@mail.ubc.ca hbouamor@cmu.edu nizar.habash@nyu.edu Abstract We present the findings and results of the Second Nuanced Arabic Dialect Identification Shared Task (NADI 2021). This Shared Task includes four subtasks: country-level Modern Standard Arabic (MSA) identification (Subtask 1.1), country-level dialect identification (Subtask 1.2), province-level MSA identification (Subtask 2.1), and province-level sub-dialect identifica- tion (Subtask 2.2). The shared task dataset cov- ers a total of 100 provinces from 21 Arab coun- Figure 1: A map of the Arab World showing thr 21 tries, collected from the Twitter domain. A total countries and 100 provinces in the NADI 2021 datasets. of 53 teams from 23 countries registered to par- Each country is coded in color different from neighbor- ing countries. Provinces within each country are coded ticipate in the tasks, thus reflecting the interest in a more intense version of the same color as the coun- of the community in this area. We received 16 try. submissions for Subtask 1.1 from five teams, 27 submissions for Subtask 1.2 from eight teams, 12 submissions for Subtask 2.1 from four teams, ple from various sub-regions, countries, or even and 13 Submissions for subtask 2.2 from four provinces within the same country, may be using teams. Arabic differently. 1 Introduction The Nuanced Arabic Dialect Identification (NADI) series of shared tasks aim at furthering Arabic is the native tongue of ∼ 400 million peo- the study and analysis of Arabic variants by provid- ple living the Arab world, a vast geographical re- ing resources and organizing classification compe- gion across Africa and Asia. Far from a single titions under standardized settings. The First Nu- monolithic language, Arabic has a wide number anced Arabic Dialect Identification (NADI 2020) of varieties. In general, Arabic could be classified Shared Task targeted 21 Arab countries and a to- into three main categories: (1) Classical Arabic, tal of 100 provinces across these countries. NADI the language of the Qur’an and early literature; (2) 2020 consisted of two subtasks: country-level di- Modern Standard Arabic (MSA), which is usually alect identification (Subtask 1) and province-level used in education and formal and pan-Arab media; detection (Subtask 2). The two subtasks depended and (3) dialectal Arabic (DA), a collection of geo- on Twitter data, making it the first shared task to tar- politically defined variants. Modern day Arabic get naturally-occurring fine-grained dialectal text is usually referred to as diglossic with a so-called at the sub-country level. The Second Nuanced Ara- ‘High’ variety used in formal settings (MSA), and bic Dialect Identification (NADI 2021) is similar a ‘Low’ variety used in everyday communication to NADI 2020 in that it also targets the same 21 (DA). DA, the presumably ‘Low’ variety, is itself Arab countries and 100 corresponding provinces a host of variants. For the current work, we focus and is based on Twitter data. However, NADI 2021 on geography as an axis of variation where peo- has four subtasks, organized into country level and
province level. For each classification level, we af- Arabic dialect identification work as further ford both MSA and DA datasets as Table 1 shows. sparked by a series of shared tasks offered as part of the VarDial workshop. These shared tasks used Variety Country Province speech broadcast transcriptions (Malmasi et al., MSA Subtask 1.1 Subtask 2.1 2016), and integrated acoustic features (Zampieri DA Subtask 1.2 Subtask 2.2 et al., 2017) and phonetic features (Zampieri et al., Table 1: NADI 2021 subtasks. 2018) extracted from raw audio. Althobaiti (2020) is a recent survey of computational work on Arabic dialects. We provided participants with a new Twitter la- beled dataset that we collected exclusively for the The Multi Arabic Dialects Application and Re- purpose of the shared task. The dataset is publicly sources (MADAR) project (Bouamor et al., 2018) available for research.1 A total of 53 teams regis- introduced finer-grained dialectal data and a lexi- tered for the shard task, of whom 8 unique teams con. The MADAR data were used for dialect iden- ended up submitting their systems for scoring. We tification at the city level (Salameh et al., 2018; allowed a maximum of five submissions per team. Obeid et al., 2019) of 25 Arab cities. An issue with We received 16 submissions for Subtask 1.1 from the MADAR data, in the context of DA identifica- five teams, 27 submissions for Subtask 1.2 from tion, is that it was commissioned and not naturally eight teams, 12 submissions for Subtask 2.1 from occurring. Several larger datasets covering 10-21 four teams, and 13 Submissions for subtask 2.2 countries were also introduced (Mubarak and Dar- from four teams. We then received seven papers, wish, 2014; Abdul-Mageed et al., 2018; Zaghouani all of which we accepted for publication. and Charfi, 2018). These datasets come from the This paper is organized as follows. We provide Twitter domain, and hence are naturally-occurring. a brief overview of the computational linguistic Several works have also focused on socio- literature on Arabic dialects in Section 2. We de- pragmatics meaning exploiting dialectal data. scribe the two subtasks and dataset in Sections 3 These include sentiment analysis (Abdul-Mageed and Section 4, respectively. And finally, we intro- et al., 2014), emotion (Alhuzali et al., 2018), duce participating teams, shared task results, and age and gender (Abbes et al., 2020), offen- a high-level description of submitted systems in sive language (Mubarak et al., 2020), and sar- Section 5. casm (Abu Farha and Magdy, 2020). Concurrent with our work, (Abdul-Mageed et al., 2020c) also 2 Related Work describe data and models at country, province, and As we explained in Section 1, Arabic has three city levels. main categories: CA, MSA, and DA. While CA The first NADI shared task, NADI 2020 (Abdul- and MSA have been studied extensively (Harrell, Mageed et al., 2020b), comprised two subtasks, 1962; Cowell, 1964; Badawi, 1973; Brustad, 2000; one focusing on 21 Arab countries exploiting Twit- Holes, 2004), DA has received more attention only ter data, and another on 100 Arab provinces from in recent years. the same 21 countries. As is explained in (Abdul- One major challenge with studying DA has been Mageed et al., 2020b), the NADI 2020 datasets rarity of resources. For this reason, most pioneer- included a small amount of non-Arabic and also a ing DA works focused on creating resources, usu- mixture of MSA and DA. For NADI 2021, we con- ally for only a small number of regions or coun- tinue to focus on 21 countries and 100 provinces. tries (Gadalla et al., 1997; Diab et al., 2010; Al- However, we breakdown the data into MSA and Sabbagh and Girju, 2012; Sadat et al., 2014; Smaı̈li DA for a stronger signal. This also gives us the et al., 2014; Jarrar et al., 2016; Khalifa et al., 2016; opportunity to study each of these two main cate- Al-Twairesh et al., 2018; El-Haj, 2020). A number gories independently. In other words, in addition to of works introducing multi-dialectal data sets and dialect and sub-dialect identification, it allows us regional level detection models followed (Zaidan to investigate the extent to which MSA itself can and Callison-Burch, 2011; Elfardy et al., 2014; be teased apart at the country and province levels. Bouamor et al., 2014; Meftouh et al., 2015). Our hope is that NADI 2021 will support exploring 1 The dataset is accessible via our GitHub at: https: variation in geographical regions that have not been //github.com/UBC-NLP/nadi. studied before.
3 Task Description While the MADAR shared task (Bouamor et al., 2019) involved prediction of a small set of cities, The NADI shared task consists of four subtasks, NADI 2020 was the first to propose automatic di- comprising two levels of classification–country and alect identification at geographical regions as small province. Each level of classification is carried out as provinces. Concurrent with NADI 2020, (Abdul- for both MSA and DA. We explain the different Mageed et al., 2020c) introduced the concept of subtasks across each classification level next. microdialects, and proposed models for identifying language varieties defined at both province and city 3.1 Country-level Classification levels. NADI 2021 follows these works, but has • Subtask 1.1: Country-level MSA. The goal one novel aspect: We introduce province-level iden- of Subtask 1.1 is to identify country level tification for MSA and DA independently (i.e., each MSA from short written sentences (tweets). variety is handled in a separate subtask). While NADI 2021 Subtask 1.1 is novel since no pre- province-level sub-dialect identification may be vious works focused on teasing apart MSA by challenging, we hypothesize province-level MSA country of origin. might be even more difficult. However, we were curious to what extent, if possible at all, a machine • Subtask 1.2: Country-level DA. Subtask 1.2 would be successful in teasing apart MSA data at is similar to Subtask 1.1, but focuses on iden- the province-level. tifying country level dialect from tweets. Sub- In addition, similar to NADI 2020, we acknowl- task 1.2 is similar to previous works that have edge that province-level classification is somewhat also taken country as their target (Mubarak related to geolocation prediction exploiting Twit- and Darwish, 2014; Abdul-Mageed et al., ter data. However, we emphasize that geolocation 2018; Zaghouani and Charfi, 2018; Bouamor prediction is performed at the level of users, rather et al., 2019; Abdul-Mageed et al., 2020b). than tweets. This makes our subtasks different from geolocation work. Another difference lies in We provided labeled data to NADI 2021 partic- the way we collect our data as we will explain in ipants with specific training (TRAIN) and devel- Section 4. Tables 11 and 12 (Appendix A) show opment (DEV) splits. Each of the 21 labels corre- the distribution of the 100 province classes in our sponding to the 21 countries is represented in both MSA and DA data splits, respectively. Importantly, TRAIN and DEV. Teams could score their models for all 4 subtasks, tweets in the TRAIN, DEV and through an online system (codalab) on the DEV TEST splits come from disjoint sets. set before the deadline. We released our TEST set of unlabeled tweets shortly before the system 3.3 Restrictions and Evaluation Metrics submission deadline. We then invited participants We follow the same general approach to managing to submit their predictions to the online scoring the shared task as our first NADI in 2020. This system housing the gold TEST set labels. Table 2 includes providing participating teams with a set of shows the distribution of the TRAIN, DEV, and restrictions that apply to all subtasks, and clear eval- TEST splits across the 21 countries. uation metrics. The purpose of our restrictions is to 3.2 Province-level Classification ensure fair comparisons and common experimen- tal conditions. In addition, similar to NADI 2020, • Subtask 2.1: Province-level MSA. The goal our data release strategy and our evaluation setup of Subtask 2.1 is to identify the specific state through the CodaLab online platform facilitated or province (henceforth, province) from which the competition management, enhanced timeliness an MSA tweet was posted. There are 100 of acquiring results upon system submission, and province labels in the data, and provinces guaranteed ultimate transparency.2 are unequally distributed among the list of Once a team registered in the shared task, we 21 countries. directly provided the registering member with the data via a private download link. We provided • Subtask 2.2: Province-level DA. Again, the data in the form of the actual tweets posted to Subtask 2.2 is similar to Subtask 2.1, but the the Twitter platform, rather than tweet IDs. This goal is identifying the province from which a 2 dialectal tweet was posted. https://codalab.org/
MSA (Subtasks 1.1 & 2.1) DA (Subtasks 1.2 & 2.2) Country Provinces Train DEV TEST Total % Train DEV TEST Total % Algeria 9 1,899 427 439 2,765 8.92 1,809 430 391 2,630 8.48 Bahrain 1 211 51 51 313 1.01 215 52 52 319 1.03 Djibouti 1 211 52 51 314 1.01 215 27 7 249 0.80 Egypt 20 4,220 1,032 989 6,241 20.13 4,283 1,041 1,051 6,375 20.56 Iraq 13 2,719 671 652 4,042 13.04 2,729 664 664 4,057 13.09 Jordan 2 422 103 102 627 2.02 429 104 105 638 2.06 Kuwait 2 422 103 102 627 2.02 429 105 106 640 2.06 Lebanon 3 633 155 141 929 3.00 644 157 120 921 2.97 Libya 6 1,266 310 307 1,883 6.07 1,286 314 316 1,916 6.18 Mauritania 1 211 52 51 314 1.01 215 53 53 321 1.04 Morocco 4 844 207 205 1,256 4.05 858 207 212 1,277 4.12 Oman 7 1,477 341 357 2,175 7.02 1,501 355 371 2,227 7.18 Palestine 2 422 102 102 626 2.02 428 104 105 637 2.05 Qatar 1 211 52 51 314 1.01 215 52 53 320 1.03 KSA 10 2,110 510 510 3,130 10.10 2,140 520 522 3,182 10.26 Somalia 2 346 63 102 511 1.65 172 49 55 276 0.89 Sudan 1 211 48 51 310 1.00 215 53 53 321 1.04 Syria 6 1,266 309 306 1,881 6.07 1,287 278 288 1,853 5.98 Tunisia 4 844 170 176 1,190 3.84 859 173 212 1,244 4.01 UAE 3 633 154 153 940 3.03 642 157 158 957 3.09 Yemen 2 422 88 102 612 1.97 429 105 106 640 2.06 Total 100 21,000 5,000 5,000 31,000 100 21,000 5,000 5,000 31,000 100 Table 2: Distribution of classes and data splits over our MSA and DA datasets for the four subtasks. . guaranteed comparison between systems exploiting report their results on the DEV set in their papers. identical data. For all four subtasks, we provided We setup four CodaLab competitions for scoring clear instructions requiring participants not to use participant systems.3 We will keep the Codalab any external data. That is, teams were required competition for each subtask live post competition, to only use the data we provided to develop their for researchers who would be interested in train- systems and no other datasets regardless how these ing models and evaluating their systems using the are acquired. For example, we requested that teams shared task TEST set. For this reason, we will do not search nor depend on any additional user- not release labels for the TEST set of any of the level information such as geolocation. To alleviate subtasks. these strict constraints and encourage creative use of diverse (machine learning) methods in system 4 Shared Task Datasets development, we provided an unlabeled dataset of 10M tweets in the form of tweet IDs. This dataset We distributed two Twitter datasets, one in MSA is in addition to our labeled TRAIN and DEV splits and another in DA. Each tweet in each of these two for the four subtasks. To facilitate acquisition of datasets has two labels, one label for country level this unlabeled dataset, we also provided a simple and another label for province level. For example, script that can be used to collect the tweets. We for the MSA dataset, the same tweet is assigned encouraged participants to use these 10M unlabeled one out of 21 country labels (Subtask 1.1) and one tweets in any way they wished. out of 100 province labels (Subtask 2.1). The same applies to DA data, where each tweet is assigned For all four subtasks, the official metric is macro- a country label (Subtask 1.2) and a province label averaged F 1 score obtained on blind test sets. (Subtask 2.2). Similar to MSA, the tagset for DA We also report performance in terms of macro- data has 21 country labels and 100 province labels. averaged precision, macro-averaged recall and ac- 3 curacy for systems submitted to each of the four Links to the CodaLab competitions are as follows: Subtask 1.1: https://competitions.codalab. subtasks. Each participating team was allowed to org/competitions/27768, Subtask 1.2: https: submit up to five runs for each subtask, and only //competitions.codalab.org/competitions/ the highest scoring run was kept as representing 27769, Subtask 2.1: https://competitions. codalab.org/competitions/27770, Sub- the team. Although official results are based only task 2.2: https://competitions.codalab.org/ on a blind TEST set, we also asked participants to competitions/27771.
In addition, as mentioned before, we made avail- to both our MSA and DA data. That is, we la- able an unlabeled dataset for optional use in any bel tweets from each user with the country and of the four subtasks. We now provide more details province from which the user consistently posted about both the labeled and unlabeled data. for the whole of the 10 months period. Although this method of label assignment is not ideal, it is 4.1 Data Collection still a reasonable approach for easing the bottleneck Similar to NADI 2020, we used the Twitter API of data annotation. For both the MSA and DA data, to crawl data from 100 provinces belonging to 21 across the two levels of classification (i.e., country Arab countries for 10 months (Jan. to Oct., 2019).4 and province), we randomly sample 21K tweets Next, we identified users who consistently and ex- for training (TRAIN), 5K tweets for development clusively tweeted from a single province during (DEV), and 5K tweets for testing (TEST). These the whole 10 month period. We crawled up to three splits come from three disjoint sets of users. 3,200 tweets from each of these users. We se- We distribute data for the four subtasks directly to lect only tweets assigned the Arabic language tag participants in the form of actual tweet text. Table 2 (ar) by Twitter. We lightly normalize tweets by shows the distribution of tweets across the data removing usernames and hyperlinks, and add white splits over the 21 countries, for all subtasks. We space between emojis. Next, we remove retweets provide the data distribution over the 100 provinces (i.e., we keep only tweets and replies). Then, we in Appendix A. More specifically, Table 11 shows use character-level string matching to remove se- the province-level distribution of tweets for MSA quences that have < 3 Arabic tokens. (Subtask 2.1) and Table 12 shows the same for DA (Subtask 2.2). We provide examples DA tweets from a number of countries representing different regions in Table 3. For each example in Table 3, we list the province it comes from. Similarly, we provide example MSA data in Table 4. Unlabeled 10M. We shared 10M Arabic tweets with participants in the form of tweet IDs. We crawled these tweets in 2019. Arabic was identified using Twitter language tag (ar). This dataset does not have any labels and we call it UNLABELED 10M. We also included in our data package released to participants a simple script to crawl these tweets. Figure 2: Distribution of tweet length (trimmed at 50) in words in NADI-2021 labeled data. Participants were free to use UNLABELED 10M for any of the four subtasks in any way they they see fits.5 We now present shared task teams and Since the Twitter language tag can be wrong results. sometimes, we apply an effective in-house lan- guage identification tool on the tweets and replies 5 Shared Task Teams & Results to exclude any non-Arabic. This helps us remove posts in Farsi (fa) and Persian (ps) which Twitter 5.1 Our Baseline Systems wrongly assigned an Arabic language tag. Finally, to tease apart MSA from DA, we use the dialect- We provide two simple baselines, Baseline I and MSA model introduced in Abdul-Mageed et al. Baseline II, for each of the four subtasks. Base- (2020a) (acc= 89.1%, F1= 88.6%). line I is based on the majority class in the TRAIN data for each subtask. It performs at F1 = 1.57% 4.2 Data Sets and accuracy = 19.78% for Subtask 1.1, F1 = 1.65% and accuracy = 21.02% for Subtask 1.2, To assign labels for the different subtasks, we use F1 = 0.02% and accuracy = 1.02% for Subtask user location as a proxy for language variety labels at both country and province levels. This applies 5 Datasets for all the subtasks and UNLABELED 10M are available at https://github.com/UBC-NLP/nadi. 4 Although we tried, we could not collect data from Co- More information about the data format can be found in the moros to cover all 22 Arab countries. accompanying README file.
Country Province Tweet Algeria Bouira úæJ Kñ» ø YK úm.' ... éJ¯ ùªJ.K ú»@P ÈAm Khenchela !, PQK @QË@ ñY®K úÍ ¼ð XAë ÕæË@ ñJ.Ê®K ÈAm Oran @ñJm'. P éK QÓ àA®«P ¼@P Egypt Alexandria AJKAJk ø YªK PY® Jë AÓ ø P Ó . Minya úÍñêÊJ®K ½ÊJj.ë úkA ù®J.Jë ø ÒÓ É¿ AK@ . Q Sohag éJËñË@ ø P úæj®K Yg YJªÓ AK@ KSA Ar-Riyad !! AÖ Þ àñJ.ªÊK àñÊm.' áK YªK ð Ash-Sharqiyah Ñê®K ÑêÖßQk úæ.K àñÒê®K AÓ éK ð@Q® ¯ ÈAg. P ÑêÖÞ @ ÈCmÌ '@ áK. AK Tabuk éJªJJ.£ èA¯ð ø Yë èP@PñË@ ÉgX @ I.J£ Morocco Marrakech-Tensift-Al-Haouz ÈðXAJK. H. Qå @ éJ .ªË@ Q» hAKP@ àA« Bð l'. P ú¯AÓ Meknes-Tafilalet . AmÌ áÊ¿@YÓ@ p úG P ñ®«@ p@X @ qJK@ àð@X@QÓ !!!! Ik . Souss-Massa-Draa ¬@QK. áJJªJ Ó AÒë úæk ¨A¿ ÑîDJ éJK @ Yg C@ Oman Ash-Sharqiyah .lÌ. ʪK úæK@ i.Êm' Dhofar éJÊ« ᮪K úæ.JÓ , à@Q¯ ùªÖÞ @ Musandam AîE PAm.' é
Country Province Tweet Algeria Biskra ÉJjÓ iJ@.. ÈCg k ¼YJ« . . AK. AJ« ð ... AK. A¢k H Qêk. ZA®ÊË@ éJ J« úΫ CK ñ£ ½KQ¢JK@ . Oran ý«ð@ ... ú¾K. @ à@ YK P@ úæK@ ... @QJ» ½KQ.g@ .. A¯A J« Q« HXP@ AÓ ð .. Ouargla ½K QK ñK. @ éÊ KAªË éJ m' Ë A¯ é£A Ë@ ñë úæ ÉÔg @ . . Egypt Faiyum Ðð QêÓ PA¯ ½«ñK. P á« ÉgQK àB@ Minya á« É¿ áÓð éÓAëð éÓB àA¢J É¿ áÓ éÓA JË@ é
Team F1 Acc Precision Recall CairoSquad 22.38(1) 35.72(1) 31.56(1) 20.66(1) Phonemer 21.79(2) 32.46(3) 30.03(3) 19.95(2) CS-UM6P 21.48(3) 33.74(2) 30.72(2) 19.70(3) Speech Translation 14.87(4) 24.32(4) 18.95(4) 13.85(4) Our Baseline II 14.15 24.76 20.01 13.21 NAYEL 12.99(5) 23.24(5) 15.09(5) 12.46(5) Our Baseline I 1.57 19.78 0.94 4.76 Table 6: Results for Subtask 1.1 (country-level MSA). The numbers in parentheses are the ranks. The table is sorted on the macro − F 1 score, the official metric. Team F1 Acc Precision Recall CairoSquad 32.26(1) 51.66(1) 36.03(1) 31.09(1) CS-UM6P 30.64(2) 49.50(2) 32.91(2) 30.34(2) Phonemer 24.29(4) 44.14(3) 30.24(3) 23.70(4) Speech Translation 21.49(5) 40.54(5) 26.75(5) 20.36(6) Arizona 21.37(6) 40.46(6) 26.32(6) 20.78(5) AraDial MJ 18.94(7) 35.94(8) 21.58(8) 18.28(7) NAYEL 18.72(8) 37.16(7) 21.61(7) 18.12(8) Our Baseline II 18.02 33.04 18.69 17.88 Our Baseline I 1.65 21.02 1.00 4.76 Table 7: Results for Subtask 1.2 (province-level MSA) Team F1 Acc Precision Recall CairoSquad 6.43(1) 6.66(1) 7.11(1) 6.71(1) Phonemer 5.49(2) 6.00(2) 6.17(2) 6.07(2) CS-UM6P 5.35(3) 5.72(3) 5.71(3) 5.75(3) NAYEL 3.51(4) 3.38(4) 4.09(4) 3.45(4) Our Baseline II 3.39 3.48 3.68 3.49 Our Baseline I 0.02 1.02 0.01 1.00 Table 8: Results for Subtask 2.1 (country-level DA). Team F1 Acc Precision Recall CairoSquad 8.60(1) 9.46(1) 9.07(1) 9.33(1) CS-UM6P 7.32(2) 7.92(2) 7.73(2) 7.95(2) NAYEL 4.55(3) 4.80(3) 4.71(3) 4.55(4) Phonemer 4.37(4) 5.32(4) 4.49(4) 5.19(3) Our Baseline II 4.08 4.18 4.54 4.22 Our Baseline I 0.02 1.06 0.01 1.00 Table 9: Results for Subtask 2.2 (province-level DA). mance with 8.60%.7 and CS-UM6P developed their system utilizing MARBERT (Abdul-Mageed et al., 2020a), a pre- 5.4 General Description of Submitted trained Transformer language model tailored to Systems Arabic dialects and the domain of social media. In Table 10, we provide a high-level descrip- Team Phonemer utilized AraBERT (Antoun et al., tion of the systems submitted to each subtask. 2020a) and AraELECTRA (Antoun et al., 2020b). For each team, we list their best score of each Team CairSquad apply adapter modules (Houlsby subtask, the features employed, and the meth- et al., 2019) and vertical attention to MARBERT ods adopted/developed. As can be seen from fine-tuning. CS-UM6P fine-tuned MARBERT on the table, the majority of the top teams have country-level and province-level jointly by multi- used Transformers. Specifically, team CairoSquad task learning. The rest of participating teams have 7 either used a type of neural networks other than The full sets of results for Subtask 1.1, 1.2, 2.1, and 2.2 are in Tables 13, 14, 15 and 15, respectively, in Appendix A. Transformers or resorted to linear machine learning
Features Techniques Word embeds Classical ML Transformer Neural nets Semi-super Linguistics Multitask Ensemble Sampling TF-IDF N-gram PMI F1 Team SUBTASK 1.1 CairoSquad 22.38 Phonemer 21.79 CS-UM6P 21.48 Speech Trans 14.87 NAYEL 12.99 SUBTASK 1.2 CairoSquad 32.26 CS-UM6P 30.64 Phonemer 24.29 Speech Trans 21.49 Arizona 21.37 AraDial MJ 18.94 NAYEL 18.72 SUBTASK 2.1 CairoSquad 6.43 Phonemer 5.49 CS-UM6P 5.35 NAYEL 3.51 SUBTASK 2.2 CairoSquad 8.60 CS-UM6P 7.32 NAYEL 4.55 Phonemer 4.37 Table 10: Summary of approaches used by participating teams. PMI: poinwise mutual information. Classical ML refers to any non-neural machine learning methods such as naive Bayes and support vector machines. The term “neural nets” refers to any model based on neural networks (e.g., FFNN, RNN, and CNN) except Transformer models. Transformer refers to neural networks based on a Transformer architecture such as BERT. The table is sorted by official metric , macro − F1 . We only list teams that submitted a description paper. “Semi-super” indicates that the model is trained with semi-supervised learning. models, usually with some form of ensembling. a continued interest in the community and calls for further work in this area. 6 Conclusion and Future Work In the future, we plan to host a third iteration of the NADI shared task that will use new datasets and We presented the findings and results of the NADI encourage novel solutions to the set of problems 2021 shared task. We described our datasets across introduced in NADI 2021. As results show all the the four subtasks and the logistics of running the fours subtasks remain very challenging, and we shared task. We also provided a panoramic descrip- hope that encouraging further solutions will help tion of the methods used by all participating teams. advance work in this area. The results show that distinguishing the language variety of short texts based on small geographical Acknowledgments regions of origin is possible, yet challenging. The total number of submissions during official evalua- We gratefully acknowledge the support of the Nat- tion (n=68 submissions from 8 unique teams), as ural Sciences and Engineering Research Council well as the number of teams who registered and of Canada, the Social Sciences Research Council acquired our datasets (n=53 unique teams) reflects of Canada, Compute Canada, and UBC Sockeye.
References Modeling of People’s Opinions, Personality, and Emotions in Social Media, pages 25–35. Ines Abbes, Wajdi Zaghouani, Omaima El-Hardlo, and Faten Ashour. 2020. Daict: A dialectal Arabic irony Badr AlKhamiss, Mohamed Gabr, Muhammed El- corpus extracted from twitter. In Proceedings of Nokrashy, and Khaled Essam. 2021. Adapting The 12th Language Resources and Evaluation Con- MARBERT for Improved Arabic Dialect Identifica- ference, pages 6265–6271. tion: Submission to the NADI 2021 Shared Task. In Proceedings of the Sixth Arabic Natural Language Muhammad Abdul-Mageed, Hassan Alhuzali, and Mo- Processing Workshop (WANLP 2021). hamed Elaraby. 2018. You tweet what you speak: A city-level dataset of Arabic dialects. In Proceedings Maha J Althobaiti. 2020. Automatic Arabic dialect of the Language Resources and Evaluation Confer- identification systems for written texts: A survey. ence (LREC), Miyazaki, Japan. arXiv preprint arXiv:2009.12622. Muhammad Abdul-Mageed, Mona Diab, and Sandra Maha J. Althobaiti. 2021. Country-level Arabic Di- Kübler. 2014. Samar: Subjectivity and sentiment alect Identification Using Small Datasets with In- analysis for Arabic social media. Computer Speech tegrated Machine Learning Techniques and Deep and Language, 28(1):20–37. Learning Models. In Proceedings of the Sixth Arabic Natural Language Processing Workshop Muhammad Abdul-Mageed, AbdelRahim Elmadany, (WANLP 2021). and El Moatez Billah Nagoudi. 2020a. ARBERT & MARBERT: Deep Bidirectional Transformers for Wissam Antoun, Fady Baly, and Hazem Hajj. 2020a. Arabic. arXiv preprint arXiv:2010.04900. Arabert: Transformer-based model for arabic lan- guage understanding. In Proceedings of the 4th Muhammad Abdul-Mageed, Chiyu Zhang, Houda Workshop on Open-Source Arabic Corpora and Pro- Bouamor, and Nizar Habash. 2020b. NADI 2020: cessing Tools, with a Shared Task on Offensive Lan- The First Nuanced Arabic Dialect Identification guage Detection, pages 9–15. Shared Task. In Proceedings of the Fifth Arabic Natural Language Processing Workshop (WANLP Wissam Antoun, Fady Baly, and Hazem Hajj. 2020b. 2020), pages 97–110, Barcelona, Spain. AraELECTRA: pre-training text discriminators for arabic language understanding. arXiv preprint Muhammad Abdul-Mageed, Chiyu Zhang, Ab- arXiv:2012.15516. delRahim Elmadany, and Lyle Ungar. 2020c. Micro-dialect identification in diaglossic and code- MS Badawi. 1973. Levels of contemporary Arabic in switched environments. In Proceedings of the Egypt. Cairo: Dâr al Ma’ârif. 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5855–5876. Houda Bouamor, Nizar Habash, and Kemal Oflazer. 2014. A multidialectal parallel corpus of Arabic. In Ibrahim Abu Farha and Walid Magdy. 2020. From Ara- Proceedings of the Language Resources and Evalu- bic Sentiment Analysis to Sarcasm Detection: The ation Conference (LREC), Reykjavik, Iceland. ArSarcasm Dataset. In Proceedings of the 4th Work- shop on Open-Source Arabic Corpora and Process- Houda Bouamor, Nizar Habash, Mohammad Salameh, ing Tools, with a Shared Task on Offensive Language Wajdi Zaghouani, Owen Rambow, Dana Abdul- Detection, pages 32–39. rahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Alexander Erdmann, and Kemal Oflazer. 2018. The Rania Al-Sabbagh and Roxana Girju. 2012. YADAC: MADAR Arabic Dialect Corpus and Lexicon. In Yet another Dialectal Arabic Corpus. In Proceed- Proceedings of the Language Resources and Eval- ings of the Eighth International Conference on uation Conference (LREC), Miyazaki, Japan. Language Resources and Evaluation (LREC-2012), pages 2882–2889. Houda Bouamor, Sabit Hassan, and Nizar Habash. 2019. The MADAR shared task on Arabic fine- Nora Al-Twairesh, Rawan Al-Matham, Nora Madi, grained dialect identification. In Proceedings of the Nada Almugren, Al-Hanouf Al-Aljmi, Shahad Al- Fourth Arabic Natural Language Processing Work- shalan, Raghad Alshalan, Nafla Alrumayyan, Shams shop, pages 199–207. Al-Manea, Sumayah Bawazeer, Nourah Al-Mutlaq, Nada Almanea, Waad Bin Huwaymil, Dalal Alqu- Kristen Brustad. 2000. The Syntax of Spoken Arabic: sair, Reem Alotaibi, Suha Al-Senaydi, and Abeer A Comparative Study of Moroccan, Egyptian, Syrian, Alfutamani. 2018. SUAR: Towards building a cor- and Kuwaiti Dialects. Georgetown University Press. pus for the Saudi dialect. In Proceedings of the Inter- national Conference on Arabic Computational Lin- Mark W. Cowell. 1964. A Reference Grammar of Syr- guistics (ACLing). ian Arabic. Georgetown University Press, Washing- ton, D.C. Hassan Alhuzali, Muhammad Abdul-Mageed, and Lyle Ungar. 2018. Enabling deep learning of emo- Mona Diab, Nizar Habash, Owen Rambow, Mohamed tion with first-person seed expressions. In Pro- Altantawy, and Yassine Benajiba. 2010. COLABA: ceedings of the Second Workshop on Computational Arabic dialect annotation and processing. In LREC
workshop on Semitic language processing, pages Dialect Identification based on Weighted Concate- 66–74. nation of TF-IDF Transformers. In Proceedings of the Sixth Arabic Natural Language Processing Work- Mahmoud El-Haj. 2020. Habibi - a multi dialect multi shop (WANLP 2021). national Arabic song lyrics corpus. In Proceed- ings of the 12th Language Resources and Evaluation Shervin Malmasi, Marcos Zampieri, Nikola Ljubešić, Conference, pages 1318–1326, Marseille, France. Preslav Nakov, Ahmed Ali, and Jörg Tiedemann. 2016. Discriminating between similar languages Abdellah El Mekki, Abdelkader El Mahdaouy, Ka- and arabic dialect identification: A report on the bil Essefar, Nabil El Mamoun, Ismail Berrada, and third dsl shared task. In Proceedings of the third Ahmed Khoumsi. 2021. CS-UM6P @ NADI’2021: workshop on NLP for similar languages, varieties BERT-based Multi-Task Model for Country and and dialects (VarDial3), pages 1–14. Province Level MSA and Dialectal Arabic Identifi- cation. In Proceedings of the Sixth Arabic Natural Karima Meftouh, Salima Harrat, Salma Jamoussi, Language Processing Workshop (WANLP 2021). Mourad Abbas, and Kamel Smaili. 2015. Machine translation experiments on padic: A parallel Arabic Heba Elfardy, Mohamed Al-Badrashiny, and Mona dialect corpus. In Proceedings of the Pacific Asia Diab. 2014. Aida: Identifying code switching in Conference on Language, Information and Compu- informal Arabic text. In Proceedings of the Con- tation. ference on Empirical Methods in Natural Language Processing (EMNLP), pages 94–101, Doha, Qatar. Hamdy Mubarak and Kareem Darwish. 2014. Using Hassan Gadalla, Hanaa Kilany, Howaida Arram, Twitter to collect a multi-dialectal corpus of Arabic. Ashraf Yacoub, Alaa El-Habashi, Amr Shalaby, In Proceedings of the Workshop for Arabic Natural Krisjanis Karins, Everett Rowson, Robert MacIn- Language Processing (WANLP), Doha, Qatar. tyre, Paul Kingsbury, David Graff, and Cynthia McLemore. 1997. CALLHOME Egyptian Arabic Hamdy Mubarak, Kareem Darwish, Walid Magdy, transcripts LDC97T19. Web Download. Philadel- Tamer Elsayed, and Hend Al-Khalifa. 2020. phia: Linguistic Data Consortium. Overview of osact4 Arabic offensive language de- tection shared task. In Proceedings of the 4th Work- R.S. Harrell. 1962. A Short Reference Grammar of Mo- shop on Open-Source Arabic Corpora and Process- roccan Arabic: With Audio CD. Georgetown clas- ing Tools, with a Shared Task on Offensive Language sics in Arabic language and linguistics. Georgetown Detection, pages 48–52. University Press. Hamada Nayel, Ahmed Hassan, Mahmoud Sobhi, and Clive Holes. 2004. Modern Arabic: Structures, Func- Ahmed El-Sawy. 2021. Data-Driven Approach for tions, and Varieties. Georgetown Classics in Ara- Arabic Dialect Identification. In Proceedings of the bic Language and Linguistics. Georgetown Univer- Sixth Arabic Natural Language Processing Work- sity Press. shop (WANLP 2021). Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Ossama Obeid, Mohammad Salameh, Houda Bouamor, Bruna Morrone, Quentin De Laroussilhe, Andrea and Nizar Habash. 2019. ADIDA: Automatic di- Gesmundo, Mona Attariyan, and Sylvain Gelly. alect identification for Arabic. In Proceedings of 2019. Parameter-efficient transfer learning for nlp. the 2019 Conference of the North American Chap- In International Conference on Machine Learning, ter of the Association for Computational Linguistics pages 2790–2799. PMLR. (Demonstrations), pages 6–11, Minneapolis, Min- nesota. Association for Computational Linguistics. Elsayed Issa. 2021. Country-level Arabic dialect identi- fication using RNNs with and without linguistic fea- Fatiha Sadat, Farnazeh Kazemi, and Atefeh Farzindar. tures. In Proceedings of the Sixth Arabic Natural 2014. Automatic identification of Arabic language Language Processing Workshop (WANLP 2021). varieties and dialects in social media. Proceedings Mustafa Jarrar, Nizar Habash, Faeq Alrimawi, Diyam of SocialNLP, page 22. Akra, and Nasser Zalmout. 2016. Curras: an anno- tated corpus for the Palestinian Arabic dialect. Lan- Mohammad Salameh, Houda Bouamor, and Nizar guage Resources and Evaluation, pages 1–31. Habash. 2018. Fine-grained Arabic dialect identi- fication. In Proceedings of the International Con- Salam Khalifa, Nizar Habash, Dana Abdulrahim, and ference on Computational Linguistics (COLING), Sara Hassan. 2016. A Large Scale Corpus of pages 1332–1344, Santa Fe, New Mexico, USA. Gulf Arabic. In Proceedings of the Language Re- sources and Evaluation Conference (LREC), Por- Kamel Smaı̈li, Mourad Abbas, Karima Meftouh, and torož, Slovenia. Salima Harrat. 2014. Building resources for Alge- rian Arabic dialects. In Proceedings of the Confer- Mohamed Lichouri, Mourad Abbas, Khaled Lounnas, ence of the International Speech Communication As- Besma Benaziz, and Aicha Zitouni. 2021. Arabic sociation (Interspeech).
Anshul Wadhawan. 2021. Dialect Identification in Nu- anced Arabic Tweets Using Farasa Segmentation and AraBERT. In Proceedings of the Sixth Arabic Natural Language Processing Workshop (WANLP 2021). Wajdi Zaghouani and Anis Charfi. 2018. ArapTweet: A Large Multi-Dialect Twitter Corpus for Gender, Age and Language Variety Identification. In Pro- ceedings of the Language Resources and Evaluation Conference (LREC), Miyazaki, Japan. Omar F Zaidan and Chris Callison-Burch. 2011. The Arabic online commentary dataset: an annotated dataset of informal Arabic with high dialectal con- tent. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Hu- man Language Technologies: short papers-Volume 2, pages 37–41. Association for Computational Lin- guistics. Marcos Zampieri, Shervin Malmasi, Nikola Ljubešić, Preslav Nakov, Ahmed Ali, Jörg Tiedemann, Yves Scherrer, and Noëmi Aepli. 2017. Findings of the vardial evaluation campaign 2017. Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Ahmed Ali, Suwon Shon, James Glass, Yves Scher- rer, Tanja Samardzic, Nikola Ljubešić, Jörg Tiede- mann, et al. 2018. Language identification and mor- phosyntactic tagging: The second vardial evaluation campaign. In Proceedings of the Fifth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2018), pages 1–17. Appendices A Data We provide the distribution Distribution of the NADI 2021 MSA data over provinces, by country (Subtask 2.1), across our our data splits in Table 11. Similarly, Table 12 shows the distribution of the DA data over provinces for all countries (Subtask 2.2) in our data splits. B Shared Task Teams & Results We provide full results for all four subtasks. Ta- ble 13 shows full results for Subtask 1.1, Table 14 for Subtask 1.2, Table 15 for Subtask 2.1, and Ta- ble 16 for Subtask 2.2.
Province Name TRAIN DEV TEST Province Name TRAIN DEV TEST ae Abu-Dhabi 211 51 51 kw Jahra 211 52 51 ae Dubai 211 52 51 lb Akkar 211 52 39 ae Ras-Al-Khaymah 211 51 51 lb North-Lebanon 211 51 51 bh Capital 211 51 51 lb South-Lebanon 211 52 51 dj Djibouti 211 52 51 ly Al-Butnan 211 52 51 dz Batna 211 52 51 ly Al-Jabal-al-Akhdar 211 52 52 dz Biskra 211 52 51 ly Benghazi 211 51 51 dz Bouira 211 12 51 ly Darnah 211 52 51 dz Béchar 211 52 31 ly Misrata 211 52 51 dz Constantine 211 51 51 ly Tripoli 211 51 51 dz El-Oued 211 52 51 ma Marrakech-Tensift-Al-Haouz 211 51 51 dz Khenchela 211 52 51 ma Meknes-Tafilalet 211 52 52 dz Oran 211 52 51 ma Souss-Massa-Draa 211 52 51 dz Ouargla 211 52 51 ma Tanger-Tetouan 211 52 51 eg Alexandria 211 51 51 mr Nouakchott 211 52 51 eg Aswan 211 52 51 om Ad-Dakhiliyah 211 51 51 eg Asyut 211 52 51 om Ad-Dhahirah 211 32 51 eg Beheira 211 52 51 om Al-Batnah 211 51 51 eg Beni-Suef 211 52 51 om Ash-Sharqiyah 211 51 51 eg Dakahlia 211 51 51 om Dhofar 211 52 51 eg Faiyum 211 52 51 om Musandam 211 52 51 eg Gharbia 211 52 51 om Muscat 211 52 51 eg Ismailia 211 52 51 ps Gaza-Strip 211 51 51 eg Kafr-el-Sheikh 211 52 20 ps West-Bank 211 51 51 eg Luxor 211 52 51 qa Ar-Rayyan 211 52 51 eg Minya 211 51 51 sa Al-Madinah 211 51 51 eg Monufia 211 52 51 sa Al-Quassim 211 51 51 eg North-Sinai 211 52 51 sa Ar-Riyad 211 51 51 eg Port-Said 211 51 51 sa Ash-Sharqiyah 211 51 51 eg Qena 211 51 51 sa Asir 211 51 51 eg Red-Sea 211 52 51 sa Ha’il 211 51 51 eg Sohag 211 51 51 sa Jizan 211 51 51 eg South-Sinai 211 51 51 sa Makkah 211 51 51 eg Suez 211 51 51 sa Najran 211 51 51 iq Al-Anbar 211 51 51 sa Tabuk 211 51 51 iq Al-Muthannia 211 52 51 sd Khartoum 211 48 51 iq An-Najaf 211 51 51 so Banaadir 211 52 51 iq Arbil 211 52 51 so Woqooyi-Galbeed 135 11 51 iq As-Sulaymaniyah 187 52 51 sy Aleppo 211 51 51 iq Babil 211 52 51 sy As-Suwayda 211 51 51 iq Baghdad 211 51 51 sy Damascus-City 211 51 51 iq Basra 211 51 51 sy Hama 211 52 51 iq Dihok 211 52 51 sy Hims 211 52 51 iq Karbala 211 52 40 sy Lattakia 211 52 51 iq Kirkuk 211 52 51 tn Ariana 211 51 51 iq Ninawa 211 52 51 tn Bizerte 211 15 51 iq Wasit 211 51 51 tn Mahdia 211 52 23 jo Aqaba 211 52 51 tn Sfax 211 52 51 jo Zarqa 211 51 51 ye Aden 211 51 51 kw Hawalli 211 51 51 ye Ibb 211 37 51 Table 11: Distribution of the NADI 2021 MSA data over provinces, by country, across our TRAIN, DEV, and TEST splits (Subtask 2.1).
Province Name TRAIN DEV TEST Province Name TRAIN DEV TEST ae Abu-Dhabi 214 52 52 kw Jahra 215 53 53 ae Dubai 214 53 53 lb Akkar 215 53 14 ae Ras-Al-Khaymah 214 52 53 lb North-Lebanon 215 52 53 bh Capital 215 52 52 lb South-Lebanon 214 52 53 dj Djibouti 215 27 7 ly Al-Butnan 214 52 53 dz Batna 215 34 10 ly Al-Jabal-al-Akhdar 215 53 53 dz Biskra 215 53 53 ly Benghazi 214 52 52 dz Bouira 215 26 53 ly Darnah 215 53 53 dz Béchar 215 53 11 ly Misrata 214 52 53 dz Constantine 215 52 53 ly Tripoli 214 52 52 dz El-Oued 215 53 52 ma Marrakech-Tensift-Al-Haouz 214 52 53 dz Khenchela 89 53 53 ma Meknes-Tafilalet 215 50 53 dz Oran 215 53 53 ma Souss-Massa-Draa 215 53 53 dz Ouargla 215 53 53 ma Tanger-Tetouan 214 52 53 eg Alexandria 214 52 52 mr Nouakchott 215 53 53 eg Aswan 214 52 52 om Ad-Dakhiliyah 214 52 53 eg Asyut 214 53 53 om Ad-Dhahirah 215 40 53 eg Beheira 214 52 52 om Al-Batnah 214 52 53 eg Beni-Suef 214 52 52 om Ash-Sharqiyah 214 52 53 eg Dakahlia 214 52 52 om Dhofar 214 53 53 eg Faiyum 214 52 53 om Musandam 215 53 53 eg Gharbia 214 52 53 om Muscat 215 53 53 eg Ismailia 214 52 53 ps Gaza-Strip 214 52 52 eg Kafr-el-Sheikh 215 52 53 ps West-Bank 214 52 53 eg Luxor 214 52 52 qa Ar-Rayyan 215 52 53 eg Minya 214 52 53 sa Al-Madinah 214 52 52 eg Monufia 215 52 53 sa Al-Quassim 214 52 52 eg North-Sinai 215 52 53 sa Ar-Riyad 214 52 52 eg Port-Said 214 52 52 sa Ash-Sharqiyah 214 52 52 eg Qena 214 52 53 sa Asir 214 52 52 eg Red-Sea 214 52 53 sa Ha’il 214 52 52 eg Sohag 214 52 52 sa Jizan 214 52 53 eg South-Sinai 214 52 53 sa Makkah 214 52 52 eg Suez 214 52 52 sa Najran 214 52 53 iq Al-Anbar 214 52 52 sa Tabuk 214 52 52 iq Al-Muthannia 215 53 53 sd Khartoum 215 53 53 iq An-Najaf 215 53 53 so Banaadir 136 40 2 iq Arbil 215 53 53 so Woqooyi-Galbeed 36 9 53 iq As-Sulaymaniyah 153 32 53 sy Aleppo 215 52 23 iq Babil 215 53 53 sy As-Suwayda 214 53 53 iq Baghdad 214 52 52 sy Damascus-City 214 52 53 iq Basra 214 52 53 sy Hama 215 53 53 iq Dihok 215 53 30 sy Hims 214 53 53 iq Karbala 215 53 53 sy Lattakia 215 15 53 iq Kirkuk 215 53 53 tn Ariana 214 52 53 iq Ninawa 215 53 53 tn Bizerte 215 16 53 iq Wasit 214 52 53 tn Mahdia 215 52 53 jo Aqaba 215 52 53 tn Sfax 215 53 53 jo Zarqa 214 52 52 ye Aden 214 52 53 kw Hawalli 214 52 53 ye Ibb 215 53 53 Table 12: Distribution of the NADI 2021 DA data over provinces, by country, across our TRAIN, DEV, and TEST splits (Subtask 2.2).
Team F1 Acc Precision Recall CairoSquad 22.38(1) 35.72(1) 31.56(3) 20.66(1) CairoSquad 21.97(2) 34.90(2) 30.01(7) 20.15(2) Phonemer 21.79(3) 32.46(6) 30.03(6) 19.95(4) Phonemer 21.66(4) 31.70(7) 28.46(8) 20.01(3) CS-UM6P 21.48(5) 33.74(4) 30.72(5) 19.70(5) CS-UM6P 20.91(6) 33.84(3) 31.16(4) 19.09(6) Phonemer 20.78(7) 32.96(5) 37.69(1) 18.42(8) CS-UM6P 19.80(8) 31.68(8) 26.69(9) 19.04(7) Speech Translation 14.87(9) 24.32(11) 18.95(14) 13.85(9) Speech Translation 14.50(10) 24.06(12) 20.24(12) 13.24(10) Speech Translation 14.48(11) 24.88(9) 22.88(10) 13.17(11) NAYEL 12.99(12) 23.24(14) 15.09(15) 12.46(12) NAYEL 11.84(13) 23.74(13) 19.42(13) 10.92(13) NAYEL 10.29(14) 24.60(10) 33.11(2) 9.83(14) NAYEL 10.13(15) 18.32(15) 11.31(16) 9.76(15) NAYEL 7.73(16) 24.06(12) 21.07(11) 8.37(16) Table 13: Full results for Subtask 1.1 (country-level MSA). The numbers in parentheses are the ranks. The table is sorted on the macro F1 score, the official metric. Team F1 Acc Precision Recall CairoSquad 32.26(1) 51.66(1) 36.03(1) 31.09(1) CairoSquad 31.04(2) 51.02(2) 35.01(2) 30.62(2) CS-UM6P 30.64(3) 49.50(4) 32.91(6) 30.34(3) CS-UM6P 30.14(4) 48.94(5) 33.20(4) 30.21(4) CS-UM6P 29.08(5) 50.30(3) 34.99(3) 29.04(5) IDC team 26.10(6) 42.70(9) 27.04(11) 25.88(6) Phonemer 24.29(7) 44.14(6) 30.24(7) 23.70(7) IDC team 24.00(8) 40.08(14) 25.57(15) 23.29(9) Phonemer 23.56(9) 43.32(8) 28.05(10) 23.34(8) Phonemer 22.72(10) 43.46(7) 28.13(9) 22.55(10) Speech Translation 21.49(11) 40.54(10) 26.75(12) 20.36(12) Arizona 21.37(12) 40.46(12) 26.32(13) 20.78(11) Speech Translation 21.14(13) 40.32(13) 25.43(16) 20.16(14) Speech Translation 21.09(14) 40.50(11) 26.29(14) 20.02(15) Arizona 20.48(15) 40.04(15) 24.09(17) 20.22(13) Arizona 19.85(16) 39.90(16) 22.89(18) 19.66(16) AraDial MJ 18.94(17) 35.94(22) 21.58(22) 18.28(17) NAYEL 18.72(18) 37.16(20) 21.61(21) 18.12(18) AraDial MJ 18.66(19) 35.54(23) 21.45(23) 18.03(19) AraDial MJ 18.09(20) 37.22(19) 21.84(20) 17.55(20) AraDial MJ 18.06(21) 38.48(17) 22.70(19) 17.39(21) IDC team 16.33(22) 29.82(25) 18.04(25) 16.10(22) NAYEL 16.31(23) 38.08(18) 32.94(5) 15.91(23) NAYEL 14.41(24) 32.78(24) 20.16(24) 14.11(24) NAYEL 13.16(25) 36.96(21) 30.00(8) 13.83(25) NAYEL 12.81(26) 26.48(26) 14.32(26) 12.66(26) AraDial MJ 4.34(27) 12.64(27) 4.33(27) 4.70(27) Table 14: Full results for Subtask 1.2 (province-level MSA).
Team F1 Acc Precision Recall CairoSquad 6.43(1) 6.66(1) 7.11(1) 6.71(1) CairoSquad 5.81(2) 6.24(2) 6.26(2) 6.33(2) Phonemer 5.49(3) 6.00(3) 6.17(3) 6.07(3) Phonemer 5.43(4) 5.96(4) 6.12(4) 6.02(4) CS-UM6P 5.35(5) 5.72(6) 5.71(7) 5.75(6) Phonemer 5.30(6) 5.84(5) 5.97(6) 5.90(5) CS-UM6P 5.12(7) 5.50(7) 5.24(8) 5.53(7) CS-UM6P 4.72(8) 5.00(8) 5.97(5) 5.02(8) NAYEL 3.51(9) 3.38(10) 4.09(9) 3.45(10) NAYEL 3.47(10) 3.56(9) 3.53(10) 3.60(9) NAYEL 3.16(11) 3.28(11) 3.38(12) 3.40(11) NAYEL 3.15(12) 3.06(12) 3.43(11) 3.07(12) Table 15: Full results for Subtask 2.1 (country-level DA). Team F1 Acc Precision Recall CairoSquad 8.60(1) 9.46(1) 9.07(1) 9.33(1) CairoSquad 7.88(2) 8.78(2) 8.27(2) 8.66(2) CS-UM6P 7.32(3) 7.92(4) 7.73(4) 7.95(3) CS-UM6P 7.29(4) 8.04(3) 8.17(3) 7.90(4) CS-UM6P 5.30(5) 6.90(5) 7.00(5) 6.82(5) NAYEL 4.55(6) 4.80(10) 4.71(6) 4.55(10) NAYEL 4.43(7) 4.88(9) 4.59(8) 4.62(9) Phonemer 4.37(8) 5.32(6) 4.49(9) 5.19(6) Phonemer 4.33(9) 5.26(7) 4.44(10) 5.14(7) Phonemer 4.23(10) 5.20(8) 4.21(11) 5.08(8) NAYEL 3.92(11) 4.12(12) 4.05(12) 4.00(12) NAYEL 3.02(12) 3.10(13) 3.19(13) 3.19(13) CS-UM6P 2.90(13) 4.20(11) 4.68(7) 4.13(11) Table 16: Full results for Subtask 2.2 (province-level DA).
You can also read